---
id: 20260710-T0-13
title: "LLM会“纠正”非裔美国人英语：激活引导技术可减少方言偏见"
title_en: "LLMs Silently Correct African American English; Activation Steering Reduces Bias"
url: https://ai.daily.yangsir.net/daily/20260710-T0-13
issue_date: 2026-07-10
publish_date: 2026-07-09T04:00:00.000Z
category: research
source_name: "arXiv cs.CL (NLP)"
source_url: https://arxiv.org/abs/2607.06845
---

# LLM会“纠正”非裔美国人英语：激活引导技术可减少方言偏见

arXiv 新研究发现，多款主流大语言模型（14B至70B参数）在理解非裔美国人英语（AAE）时，会系统性地将其“纠正”为标准英语，造成信息误读和方言歧视。研究团队提出了一种基于激活引导（Activation Steering）的干预方法，能在不影响整体性能的情况下，有效减少模型对AAE的偏见。该研究关注AI公平性，对服务多元用户群体的开发者具有参考价值。

## English Version

**LLMs Silently Correct African American English; Activation Steering Reduces Bias**

A new arXiv study reveals that major LLMs (14B-70B) systematically 'correct' African American English (AAE) to standard English, leading to misinterpretation and dialect bias. The researchers propose an activation steering intervention that reduces this bias without harming overall model performance. The work highlights AI fairness issues and offers a practical technique for developers serving diverse user populations.

---

**来源**：[arXiv cs.CL (NLP)](https://arxiv.org/abs/2607.06845)

**详情页**：https://ai.daily.yangsir.net/daily/20260710-T0-13

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*